Skip to content

thinking_about_TA.md

Latest commit

 

History

History
95 lines (71 loc) · 5.21 KB

thinking_about_TA.md

File metadata and controls

95 lines (71 loc) · 5.21 KB
Raw

About Technical Indicators

There are three types of technical indicators

  1. Time series indicators. Must be sorted by time, and then calculated. It is difficult to handle null in the middle.
  2. Cross-sectional indicators. No order requirement, the whole row is calculated, and there may be a null, which needs to be compatible.
  3. Single element indicators. Can be single-column, multi-column, and can be calculated in any way.

In terms of calculation, time series indicators have one more dimension than cross-sectional indicators. For example

  1. cs_rank is a cross-sectional sort

  2. ts_rank is to calculate cs_rank in a rolling time window, take [-1] each time, and then concatenate them

  3. cs_zscore first aggregates to calculate mean and std, then broadcast mean and std, and perform a one-dimensional calculation with x

So common technical indicators are generally the flexible application of aggregation and broadcasting

Since ts_ indicators are based on cs_, theoretically rolling can be used directly, but in practice, it is generally not used in this way. cs_ is generally rewritten in cython, numba, etc., and if you mix ts_ and cs_, ts_ will make thousands of cs_ call. So it is common to put rolling operations into cython, numba, etc. as well.

Evolve of Our TA-Lib Wrappers

  1. Expr.map_batches can be used to call third-party libraries, such as TA-Lib, bottleneck. But because of the input and output format requirements, you need to wrap the third-party API with a function.

    • Both input and output can only be one column. If you want to support multiple columns, you need to convert them to pl.Struct. After that, you need to use unnest to split pl.Struct.
    • The output must be pl.Series

  2. Start to use register_expr_namespace to simplify the code

    • Implementation helper.py
    • Usage demo demo_ta1.py
    • Pros: Easy to use
    • Cons:
      • The member function call mode is not convenient for inputting into genetic algorithms for factor mining
      • __getattribute__ dynamic method call is very flexible, but loses IDE support.

  3. Prefix expression. Convert all member functions into formulas

    • Implementation wrapper.py
    • Usage demo demo_ta2.py
    • Pros: Can be input into our implementation of genetic algorithms
    • Cons: __getattribute__ dynamic method call is very flexible, but loses IDE support.

  4. Code generation.

对技术指标的再思考

用使用方式来说，指标分为三种

  1. 时序指标。必须按时间排序，然后计算，中段出现null比较难处理
  2. 截面指标。对顺序无要求，整行计算即可，数据会有null，需要能兼容
  3. 单元素指标。可单列、多列，可任意方式计算

从计算原理上来说，时序指标比截面指标多了一个计算维度。例如

  1. cs_rank是横截面排序
  2. 而ts_rank是对数据滚动时间窗口计算cs_rank,每次取[-1]，然后拼接起来

又例如

  1. cs_zscore先聚合计算mean和std
  2. 然后将mean与std广播,与x进行一维计算

所以常见的技术指标一般是聚合与广播的灵活应用

由于ts_指标基于cs_,理论上直接rolling即可，但在实践中一般不这么用，因为为了求快，cs_一般是更底层的语言编写，如cython、numba等， 本来cs_版python与底层只交互一次，而ts_调用cs_版会导致交互千万次，性能极低 一般是把rolling操作也放在底层，在底层循环调用cs_底层版

TA-Lib封装的演化

  1. Expr.map_batches可以实现调用第三方库，如TA-Lib, bottleneck。但因为对输入与输出格式有要求，所以还需要用函数对第三方API封装一下。
    • 输入输出都只能是一列，如要支持多列需转换成pl.Struct。事后pl.Struct要拆分需使用unnest
    • 输出必须是pl.Series
  2. 参数多，代码长。开始使用register_expr_namespace来简化代码
    • 实现代码helper.py
    • 使用演示demo_ta1.py
    • 优点：使用简单
    • 不足：成员函数调用模式不便于输入到遗传算法中进行因子挖掘
    • 不足：__getattribute__动态方法调用非常灵活，但失去了IDE智能提示
  3. 前缀表达式。将所有的成员函数都转换成公式
    • 实现代码wrapper.py
    • 使用演示demo_ta2.py
    • 优点：可以输入到遗传算法
    • 不足：__getattribute__动态方法调用非常灵活，但失去了IDE智能提示
  4. 代码自动生成。